Meta is testing its first internally designed artificial intelligence (AI) training chip, a strategic move aimed at reducing dependency on external suppliers like Nvidia and streamlining costs as the company doubles down on AI-driven growth, according to two sources familiar with the matter, as reported by Reuters.
The social media giant—parent company of Facebook, Instagram, and WhatsApp—has initiated a limited deployment of the new silicon and plans to scale production if the trials prove successful. The project, developed in collaboration with Taiwan Semiconductor Manufacturing Company (TSMC), represents a critical step in Meta's multiyear effort to overhaul its infrastructure and rein in soaring expenses tied to AI investments.
Cost Efficiency and Specialized Design
The new chip, part of Meta's Meta Training and Inference Accelerator (MTIA) program, is engineered as a dedicated AI accelerator, optimizing power efficiency by focusing exclusively on AI tasks. This contrasts with traditional GPUs, which handle broader computing workloads. A successful rollout could help Meta tackle its mammoth infrastructure costs, with the company projecting 2025 capital expenditures of up to $65 billion, largely allocated to AI development.
Sources indicate Meta recently completed the initial "tape-out" phase—finalizing the chip design and beginning factory production—a high-stakes milestone that consumes millions of dollars and months of work. Earlier iterations of the MTIA program faced hurdles, including the cancellation of a chip at a similar developmental stage. However, Meta began deploying its first MTIA inference chip in 2023 to power recommendation algorithms for Facebook and Instagram feeds.
Long-Term AI Ambitions
Meta executives aim to transition to proprietary chips for training AI models by 2026, starting with recommendation systems before expanding to generative AI products like its Meta AI chatbot. Training, which involves feeding vast datasets to AI systems, demands immense computational power—an area dominated by Nvidia's hardware.
"We're exploring how to handle training for recommender systems and eventually generative AI," Chief Product Officer Chris Cox said at a recent Morgan Stanley conference. The shift underscores Meta's broader strategy to control costs while scaling AI capabilities.
Meta and TSMC declined to comment on the project. If successful, the chip could mark a turning point in the tech industry's race to develop alternatives to Nvidia's market-leading solutions.